-
Notifications
You must be signed in to change notification settings - Fork 414
Multi-label Training got train-logloss:-nan #195
Comments
Please try to set the |
Thanks. Trying now. |
round 0:[ 1656] 7834 sec elapsed[1] train-logloss:0.0965824 train-error:0.964808 I am still getting the -nan train-logloss.... |
Hi,
I noticed that cxxnet provide an easy way to do the multi-label classification. But I met some issues duing the training.
Following are some parts of my config
layer[23->24] = fullc:fc8_fine_tune
nhidden = 256
layer[24->24] = multi_logistic
target = label
You can see I have 256 labels for each image so I set fc8_fine_tune out put as 256. The final layer is the loss layer and I am using the elementwise logistic loss function.
evaluation metric
metric = logloss
metric = error
I have two evaluation metric, one 'logloss' and one 'error'. I got following output during the training
initializing end, start working
continuing from round 0[1]
round 0:[ 1655] 7819 sec elapsed[1] train-logloss:-nan train-error:0.963257
round 1:[ 180]......
I got the '-nan' of training logloss. I am not sure whether it will become some real numbers or not with more training round. But can I have some suggestions about why this happen and how to avoid this? My learning parameters are set as following:
updater = sgd
momentum = 0.9
wmat:eta = 0.001
wmat:wd = 0.0005
bias:eta = 0.002
bias:wd = 0.0005
The text was updated successfully, but these errors were encountered: